Overview
dsh-vision-recognizer
Adaptive image routing for DeepSeek Harness: pass images directly to native multimodal models and transcribe them only for text-only models.
README / EN
Package documentation
Registry summary
Adaptive image routing for DeepSeek Harness: pass images directly to native multimodal models and transcribe them only for text-only models.
dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.
Read the full README on GitHubLIMITATIONS
Known limitations
Capability lookup and prepared target dispatch are separate public operations in rc8. A target adapter replaced by HMR in that tiny interval can race the routing decision. Nested target delegation also enters the `llm/stream` waterfall a second time, and DSH may strip provider-private replay metadata when wrapper and target adapters differ. Ordinary text/image history is preserved; provider-specific replay signatures may lose their optimization or fidelity until DSH exposes an atomic delegation handle.